Hybrid high-precision autofocus method based on deep learning and variable-step hill-climbing

In traditional autofocus algorithms, the determination of the optimal focus position relies on identifying the lens location that yields the maximum image sharpness. However, during the initial search stage, the difficulty in selecting an appropriate step size often causes the algorithm to be trapped in local optima, and under complex lighting and scene conditions, it further suffers from time-consuming iterative computations, leading to prolonged processing time. Meanwhile, current deep learning-based autofocus methods still face challenges in achieving high focusing accuracy. To address these challenges, we propose a hybrid high-precision autofocus algorithm that combines an improved deep learning network with a variable-step hill-climbing strategy. Specifically, the convolutional block attention module (CBAM) and efficient channel attention (ECA) mechanisms are integrated into ShuffleNetV2 to enhance multi-level feature extraction. The classification layer is replaced with a three-layer fully connected structure to directly regress the defocus distance. The predicted value is then used to guide the variable-step local search, thus achieving final accurate localization. Experimental results demonstrate that the proposed method reduces the average focusing error by 47% to 91% and decreases processing time by 63% to 88%. In addition, robustness is improved, with a 26% to 81% reduction in the standard deviation of the average focusing errors. Together, these improvements offer an efficient and reliable autofocus solution for complex imaging scenarios.